Reflections on Dynamic Languages and Parallelism

Post on 24-Feb-2016

40 views 0 download

description

Reflections on Dynamic Languages and Parallelism. David Padua University of Illinois at Urbana-Champaign. Parallel Programming. Parallel programming is Sometimes for Productivity - PowerPoint PPT Presentation

Transcript of Reflections on Dynamic Languages and Parallelism

1

Reflections on Dynamic Languagesand Parallelism

David PaduaUniversity of Illinois at Urbana-Champaign

2

Parallel Programming Parallel programming is

Sometimes for Productivity Because some problems are more naturally solved in parallel.

For example, some simulations, reactive programs. Most often for Performance

Serial machines are not powerful enough For scalability across machine generations. Scalability more

important than absolute performance for microprocessor industry.

Parallel programming can be Implicit – Library/Runtime/compiler Explicit – Threading, multiprocessing, parallel loops

Shared-memory Distributed memory

3

Dynamic Languages Dynamic languages are for

Productivity. They “Make programmers super productive”.

Not performance DLs are typically slow.

10-100 (1000 ?) times slower than corresponding C or Fortran

Sufficiently fast for many problems and excellent for prototyping in all cases But must manually rewrite prototype if

performance is needed.

4

Parallel Programming with Dynamic Languages Not always accepted by the DL community

Hearsay: javascript designers are unwilling to add parallel extensions.

Some in the python community prefer not to remove GIL – serial computing simplifies matters.

Not (always) great for performance Not much of an effort is made for a highly

efficient, effective form of parallelism. For example, Python’s GIL and its implementation. In MATLAB, programmer controlled communication from

desktop to worker.

5

Parallel Programming with Dynamic Languages (cont.) Not (always) great to facilitate expressing

parallelism (productivity) In some cases (e.g. MATLAB) parallel programing

constructs were not part of the language at the beginning.

Sharing of data not always possible. Python it seems that arrays can be shared between

processes, but not other classes of data. In MATLAB, there is no shared memory.

Message passing is the preferred form of communication. Process to process in the case of Python. Client to worker in the case of MATLAB MATLAB’s parfor has complex design rules

6

Why Parallel Dynamic Language Programs ? There are reasons to improve the current situation Parallelism might be necessary for dynamic

languages to have a future in the multicore era. Lack of parallelism would mean no performance

improvement across machine generations. DLs are not totally performance oblivious. They are

enabled by very powerful machines. When parallelism is explicit

For some problems it helps productivity Enable prototyping of high-performing parallel codes. Super productive parallel programming ?

Can parallelism be used to close the performance gap with conventional languages ?

7

Detour. The real answer But, if you want performance, you don’t need

parallelism, all you need is a little

8

MaJIC

9

MaJIC Results

10

How to introduce parallelism1. Autoparallelization

Parallelism is automatic via compiler/interpreter Perfect productivity But the technology does not work in al cases. Not

even for scientific programs. Next slide shows an simple experiment on

vectorization Three compilers and a few simple loops. Technology is not there not even for vectorization.

11

How to introduce parallelism1. Autoparallelization (cont.)

12

How to introduce parallelism1. Autoparallelization (cont.)

Would dynamic compilation improve the situation ?

NO

13

How to introduce parallelism2. Libraries of parallel kernels Again, programmer does not need to do

anything Again, perfect from productivity point of view. This is the performance model of MATLAB.

Good performance if most of the computation were represented in terms of kernels.

Parallelism if most of the computation were represented in terms of parallel kernels.

But not all programs can be written in terms of library routines.

14

How to introduce parallelism3. (Data) Parallel Operators

Data parallel programs have good properties Can be analyzed using sequential semantics Parallelism is encapsulated Can be used to enforce determinacy For scientific codes, array notation produces highly compact

and (sometimes) readable programs. Array notation introduced before parallelism (APL ca. 1960).

Recent (Re)Emergence of data parallel languages (e.g. Ct, ) But they are explicitly parallel

More complex program development than their sequential counterpart.

Not all forms of parallelism can be nicely represented Pipelining General task

15

Blocking/tiling crucial for locality and parallel programming.

Our approach makes tiles first class objects. Referenced explicitly. Manipulated using array operations such as

reductions, gather, etc..

Joint work with IBM Research.G. Bikshandi, J. Guo, D. Hoeflinger, G. Almasi, B. Fraguela, M. Garzarán, D. Padua, and C. von Praun. Programming for Parallelism and Locality with Hierarchically Tiled. PPoPP, March 2006.

3. (Data) Parallel OperatorsExtending MATLAB: Hierarchically Tiled Arrays

16

2 X 2 tilesmap to distinct modulesof a cluster

4 X 4 tilesUse to enhance locality on L1-cache

3. (Data) Parallel OperatorsExtending MATLAB: Hierarchically Tiled Arrays

17

h{1,1:2}

h{2,1}

h{2,1}(1,2)

tiles

3. (Data) Parallel OperatorsExtending MATLAB: Hierarchically Tiled Arrays

18

for I=1:q:n for J=1:q:n for K=1:q:n for i=I:I+q-1 for j=J:J+q-1 for k=K:K+q-1 C(i,j)=C(i,j)+A(i,k)*B(k,j); end end end end endend

for i=1:m for j=1:m for k=1:m C{i,j}=C{i,j}+A{i,k}*B{k,j}; end endend

3. (Data) Parallel OperatorsSequential MMM in MATLAB with HTAs

19

T2{:,:}

B

T1{:,:}A matmul

function C = summa (A, B, C)for k=1:m

T1 = repmat(A{:, k}, 1, m); T2 = repmat(B{k, :}, m, 1); C = C + matmul(T1{:,:} ,T2 {:,:}); end

repmat

repmat

broadcast

parallel computation

3. (Data) Parallel OperatorsParallel MMM in MATLAB with HTAs

20

How to introduce parallelism4. General mechanisms

This means Fork, join Parallel loops Synchronization, semaphores, monitors

Already in many languages. May need improvement, but no conceptual difficulty.

Maximize flexibility/maximize complexity But, Good bye super productivity !

Race conditions Tuning

21

Conclusions

DLs of the future are likely to accommodate parallelism better than they do today.

There are several possible approaches to introduce parallelism, but none is perfect.

When (if?) parallelism becomes the norm: Performance will have a more important role in DL

programming Therefore, at least for some classes of problems, super

productivity will suffer. Advances in compilers, libraries, and language

extensions will help recover some of the lost ground